Analysis of gradient descent methods with non-diminishing, bounded errors